costs and limits
How projects, concurrency, monthly allowances, and per-goal cost caps protect you.
Every plan has project limits, concurrent-agent limits, a monthly platform allowance, and per-goal cost caps. This page explains what they mean in practice.
What counts as a task
One task = one plain-English request that ends in a PR or a failure log.
- Retries on the same task are free.
- A task that runs three internal iterations to land a PR still counts as one.
- A task you cancel before the PR opens still records its model usage, but it does not create a separate task quota problem because public plans are not capped by task count.
Monthly allowance by plan
| Plan | Monthly platform allowance |
|---|---|
| Free | $100/mo of the included Kimi model |
| BYOK | $100/mo Kimi allowance + your own keys |
| Solo | $120/mo total platform allowance |
| Fleet | $180/mo total platform allowance |
| Pro | $300/mo total platform allowance |
Free is permanent, not a trial: you get up to $100/mo of the included Kimi model on us. When your plan allowance is exhausted, platform-funded inference pauses until prepaid credits are available, you switch eligible work to BYOK, or the next cycle starts.
Prepaid credits are only used within the spend limit you configure. FlareCode does not silently run up uncapped overages.
Project limits
Repos are unlimited, but each plan caps how many projects you can keep active. A project usually maps to one repo or scratch workspace and contains one or more agents.
| Plan | Projects |
|---|---|
| Free | 5 |
| BYOK | 5 |
| Solo | 10 |
| Fleet | 25 |
| Pro | 100 |
Deleting or archiving a project frees a slot immediately. When you're at the cap, creating another project asks you to delete one or upgrade.
Concurrent agents
Concurrent agents are agents actively running work at the same time.
| Plan | Concurrent agents |
|---|---|
| Free | 5 |
| BYOK | 5 |
| Solo | 20 |
| Fleet | 50 |
| Pro | 200 |
Cost cap per task
Every goal carries a hard maxCostUsd. Defaults vary by plan. If the agent exceeds it, we kill the goal immediately, post the failure log, and do not open a PR.
A runaway agent that bypasses the cap is a P0 incident on our end. If you ever see one, email us.
How we keep your bundled cost low
Whole-repo context gets cached on first use and reused across tasks for the same repo. This drops the marginal task cost dramatically. The pricing math only works because of caching — which is why it's an engineering must-have, not a nice-to-have.
What's not metered
- Failed tasks (we eat the cost).
- Setup time (sandbox boot, clone, install).
- Idle time while an agent is paused.
What is metered
- Platform-funded model inference for the task itself (input + output tokens).
- BYOK model inference is billed by your provider, not marked up by FlareCode.